Add v81 (forge host-conv, WER 0) + README v81 section

README.md

@@ -12,14 +12,58 @@ language:
 pipeline_tag: automatic-speech-recognition
 ---
 
-# whisper-base — QHexRT NPU bundle (Hexagon v79)
+# whisper-base — QHexRT NPU bundle (Hexagon v79 + v81)
+
+Precompiled **Whisper-base** ASR for the **QHexRT** runtime on Qualcomm Hexagon. Two arch-pinned bundles in
+this repo (the host pipeline — log-mel feature extraction, the decode loop, detok — is QHexRT's own; the
+encoder + decoder are prepared QNN context binaries):
+
+| arch | dir | device | source | status |
+|------|-----|--------|--------|--------|
+| **v81** | `v81/` | SM8850 / Snapdragon 8 Elite (e.g. Galaxy S26) | forge (host-conv) | **device-validated — WER 0** vs HF |
+| **v79** | `v79/` | SM8750 / Snapdragon 8 Elite (e.g. Galaxy S25) | Qualcomm AI Hub | device-validated — matches HF |
+
+## v81 (`v81/`) — forge host-conv bundle
+
+Device-validated on **SM8850 / v81** (QAIRT 2.47): `qhx_asr` transcribes a LibriSpeech clip →
+*"Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel."* — token-for-token
+equal to the HF `openai/whisper-base` reference (**WER 0**), 27 tokens in **~1.4 s**; on-device encoder
+cross-KV cosine **0.99995** vs host.
+
+The encoder graph starts at **post-conv features**: the two conv1d+GELU stem layers are HTP-hostile in-graph
+(they collapse the transcript to empty), so they run **on the host** in `asr_transcribe` from the baked
+`whisper_conv_stem.bin`. Per-stage on-device localization proved the rest of the encoder is bit-precise on v81.
+
+| file | what |
+|------|------|
+| `whisper-base.json` | QHexRT manifest (ASR family, `asr_transcribe` plan; `conv_weights` + `n_mels`) |
+| `whisperbase_enc_f16.bin` | encoder context binary — **post-conv** features → per-layer cross-attn KV |
+| `whisperbase_dec_f16.bin` | autoregressive decoder (causal self-attn + cross-attn, in-graph int32 embed + tied lm-head) |
+| `whisper_conv_stem.bin` | host-side conv-stem weights `[c1w,c1b,c2w,c2b,pos]` (the host computes the conv) |
+| `whisper_base_mel_filters.bin` | log-mel filterbank `[201,80]` (host feature extraction) |
+| `tokenizer.json` | Whisper multilingual tokenizer (vocab 51865) |
+
+```bash
+hf download runanywhere/whisper_base_HNPU --include "v81/*" --local-dir whisper_base_HNPU
+D=/data/local/tmp/whisper && adb push whisper_base_HNPU/v81/* $D/
+# + the QAIRT 2.47 aarch64 libs (libQnnHtp.so, libQnnSystem.so) + the v81 HTP skel + qhx_asr into $D
+adb shell "cd $D && export ADSP_LIBRARY_PATH='$D;/vendor/dsp/cdsp'; LD_LIBRARY_PATH=. \
+  ./qhx_asr whisper-base.json libQnnHtp.so libQnnSystem.so . <audio16k>.wav"
+# -> TRANSCRIPT: ...
+```
+
+> **Runtime caveat (v81):** this bundle needs a `qhx_asr` **built from the forge QHexRT** (the host-side-conv
+> `asr_transcribe`). That host-op branches on the encoder input — `encoder_features` → host conv (this bundle);
+> `input_features` → the AI-Hub in-graph-conv path — so the **same `qhx_asr` also runs the v79 bundle below**.
+
+## v79 (`v79/`) — Qualcomm AI Hub bundle
 
 Precompiled **Whisper-base** ASR for the **QHexRT** runtime on Qualcomm Hexagon **v79**
 (Snapdragon 8 Elite / SM8750, e.g. Galaxy S25). Encoder + decoder are Qualcomm AI Hub
 `qnn_context_binary` (float/fp16) graphs; the host pipeline (log-mel, decode loop, detok) is QHexRT's
 own. Device-validated: transcription matches the HF `openai/whisper-base` reference exactly.
 
-## Contents (`v79/`)
+### Contents (`v79/`)
 | file | what |
 |------|------|
 | `whisper-base.json` | QHexRT manifest (ASR family, `asr_transcribe` plan) |
@@ -28,7 +72,7 @@ own. Device-validated: transcription matches the HF `openai/whisper-base` refere
 | `whisper_base_mel_filters.bin` | HF mel filter bank `[201,80]` f32 (host log-mel) |
 | `tokenizer.json` | Whisper multilingual tokenizer (vocab 51865) |
 
-## Run (QHexRT CLI)
+### Run (QHexRT CLI)
 ```bash
 huggingface-cli download runanywhere/whisper_base_HNPU --local-dir whisper_base_HNPU
 # QNN libs come from the QAIRT SDK (lib/aarch64-android) + the v79 HTP skel; push them next to qhx_asr.
@@ -39,7 +83,7 @@ adb shell "cd /data/local/tmp/wq && LD_LIBRARY_PATH=. ADSP_LIBRARY_PATH=. \
 ```
 Audio: mono WAV (PCM16 or float32); resampled to 16 kHz host-side. Clips ≤ 30 s.
 
-## Notes
+### Notes
 - Arch: **v79** only (context binaries are dsp-arch-pinned). Other arches = re-export from AI Hub.
 - No custom op-package needed — these are pure-native AI Hub graphs.
 - Source model: `openai/whisper-base`, compiled via Qualcomm AI Hub for `qualcomm-snapdragon-8-elite-for-galaxy`.

v81/whisper-base.json

@@ -0,0 +1,46 @@
+{
+  "schema_version": 1,
+  "_comment": "Whisper-base ASR (encoder-decoder). Host log-mel feature extraction feeds the encoder graph; the autoregressive decoder (causal self-attention + cross-attention to the cached encoder states, learned positional embeddings, tied lm-head) emits tokens, detok'd on the host. Gated on WER vs HF Whisper.",
+  "model": {
+    "name": "whisper-base",
+    "family": "asr",
+    "dsp_arch": "v81",
+    "tokenizer_pre": "whisper"
+  },
+  "params": {
+    "hidden": 512,
+    "vocab": 51865,
+    "n_layers": 6,
+    "max_ctx": 200,
+    "kv_dim": 512,
+    "head_dim": 64,
+    "rope_theta": 10000.0,
+    "eos_token_id": 50257
+  },
+  "artifacts": {
+    "contexts": {
+      "encoder": {
+        "bin": "whisperbase_enc_f16.bin"
+      },
+      "decoder": {
+        "bin": "whisperbase_dec_f16.bin"
+      }
+    },
+    "tokenizer": "tokenizer.json"
+  },
+  "plan": {
+    "steps": [
+      {
+        "host": "asr_transcribe",
+        "params": {
+          "encoder": "whisperbase_enc_f16",
+          "decoder": "whisperbase_dec_f16",
+          "sot": 50258,
+          "mel_filters": "whisper_base_mel_filters.bin",
+          "conv_weights": "whisper_conv_stem.bin",
+          "n_mels": 80
+        }
+      }
+    ]
+  }
+}

v81/whisper_base_mel_filters.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:039ce818842793b3547f84ca19dbf32488538b14016d3cb259342ed6667903c4
+size 64320

v81/whisper_conv_stem.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab32f7839de3b8db4e4eed32b743f774fce82b00b0a2a414708272a0971a2f26
+size 6713344

v81/whisperbase_dec_f16.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e5aa7eeeb643978e7017fb4dcc962b65cf1cf648858da2deed2a8a927e93f94
+size 152477696

v81/whisperbase_enc_f16.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0c2fb3aea60453149125b636690930578c5767408c3c593ae52a748b3a38d9f
+size 58335232